BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension

We present BART, a denoising autoencoder for pretraining sequence-to-sequence. BART is trained by (1) corrupting text with an arbitrary noising, and (2) learning a model to reconstruct the original text. It uses a Tranformer-based neural machine translation architecture which, its simplicity, can be seen as generalizing BERT (due to the encoder), GPT (with the left-to-right decoder), and many other recent pretraining schemes. We evaluate a number of noising approaches, the best performance by both randomly shuffling the order of the sentences and using a novel in-filling scheme, where spans of text are with a single mask token. BART is particularly effective when fine for text generation but also works well for comprehension tasks. It the performance of RoBERTa with comparable training resources on GLUE SQuAD, achieves new state-of-the-art results on a range of abstractive, question answering, and summarization tasks, with gains of up to 6. BART also provides a 1. 1 BLEU increase over a back-translation system machine translation, with only target language pretraining. We also report experiments that replicate other pretraining schemes within the BART, to better measure which factors most influence end-task performance.